⚡ electrical engineering

Incremental Learning-Based Open-Set Classification of Unknown UAVs via RF Signal Semantics

This paper proposes a unified incremental open-set learning framework that leverages RF signal semantics to detect unknown UAVs, automatically discover new categories via adaptive clustering, and incrementally integrate them while mitigating catastrophic forgetting, thereby enabling robust recognition in dynamic real-world airspace.

Original authors: Julie Liu, Irshad A. Meer, Cicek Cavdar, Mustafa Ozger

Published 2026-03-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Julie Liu, Irshad A. Meer, Cicek Cavdar, Mustafa Ozger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard at a busy airport. Your job is to spot drones in the sky.

The Old Way (Closed-Set Learning):
In the past, security guards were trained with a photo album containing pictures of exactly 18 specific drone models. If a drone flew by, the guard would check the album.

If the drone matched a photo, they'd say, "That's a DJI Mini!"
If a new drone flew by (one not in the album), the guard would still force a match. They might squint and say, "It looks a bit like the DJI Mini, so it must be that!" This is dangerous because they are overconfident about the wrong answer. They can't say, "I don't know what this is."

The New Way (This Paper's Solution):
The researchers (Julie Liu and her team) built a "Smart Guard" system that can handle the real world, where new drones appear constantly. They call this Incremental Open-Set Learning.

Here is how their system works, broken down into three simple steps using analogies:

1. The "Translator" (Semantic Embedding)

First, the system listens to the radio "whispers" (RF signals) of the drones. These signals are messy and complex.

The Analogy: Imagine the system is a translator who converts a foreign language (radio waves) into a universal "idea language." Instead of hearing "beep-boop," the translator says, "This sounds like a fast, small bird," or "This sounds like a slow, heavy truck."
How it works: It uses advanced math (like a mix of CNNs and Transformers) to turn the radio noise into a clean, organized map where similar drones cluster together and different ones stay far apart.

2. The "Bouncer" (Open-Set Recognition)

Once the signal is translated, the system acts like a bouncer at a club.

The Analogy: The bouncer has a list of VIPs (the 18 known drones). He checks the ID of everyone who approaches.
- If the ID matches the list perfectly, they get in.
- If the ID is weird or doesn't match anyone on the list, the bouncer doesn't guess. He says, "You are unknown," and sends them to a holding area.
The Magic: Unlike old systems, this bouncer admits when he doesn't know someone. He uses a "distance rule" (Mahalanobis distance) to measure how far a new signal is from the known VIPs. If it's too far, it gets rejected.

3. The "Detective & Archivist" (Clustering & Incremental Learning)

This is where the system gets really smart. It doesn't just throw the "unknowns" away; it learns from them.

The Detective (Clustering): The rejected "unknown" drones are put in a waiting room. The system acts like a detective, looking at their "ideas" (features) and grouping them.
- Analogy: If 10 unknown drones all sound like "fast, small birds," the detective groups them together and says, "Okay, we have a new category here: Fast Bird Type A."
- The system is smart enough to pick the best way to group them (using a mix of math tools called K-Means and GMM) so it doesn't make mistakes.
The Archivist (Incremental Learning): Now the system has a new category. It needs to learn it without forgetting the old ones.
- The Problem: Usually, when you learn something new, your brain forgets the old stuff (Catastrophic Forgetting).
- The Solution: The system keeps a tiny, "memory-bounded" photo album of the old VIPs (Replay Memory). When it learns the new "Fast Bird" category, it briefly reviews the old photos to make sure it doesn't forget them.
- The Result: It updates its brain to recognize the new drone, keeps the old knowledge safe, and requires very little storage space.

Why Does This Matter?

Real-World Safety: In the real sky, new drone models appear every day. This system doesn't crash or guess wrong; it says "I don't know" and then learns the new type.
Efficiency: It doesn't need to retrain the whole system from scratch every time a new drone shows up. It just adds a small "patch" to its knowledge.
Low Cost: It works with very little memory, meaning it could run on small devices in the field, not just massive supercomputers.

In a Nutshell:
This paper teaches a computer to be a security guard that admits when it doesn't know a drone, figures out what kind of new drone it is, and adds it to its memory book without forgetting the old ones. It turns a rigid, "all-or-nothing" system into a flexible, lifelong learner.

1. Problem Statement

The proliferation of Unmanned Aerial Vehicles (UAVs) necessitates reliable Radio Frequency (RF) identification systems. However, existing RF-based recognition methods suffer from two critical limitations in real-world deployment:

Closed-Set Assumption: Traditional models assume all UAV types are known during training. In dynamic airspace, new or modified UAVs frequently emerge, causing these models to produce overconfident misclassifications of unknown signals.
Catastrophic Forgetting: While Open-Set Recognition (OSR) can detect unknowns, it typically rejects them without learning. Incremental Learning (IL) can adapt to new classes, but RF signals are complex-valued, environment-dependent, and non-stationary, making stable adaptation difficult without retraining from scratch or suffering from catastrophic forgetting (loss of knowledge of old classes).

Objective: To develop a unified framework that performs Open-Set Recognition (detecting unknowns), Novel Class Discovery (clustering unknowns into new categories), and Incremental Learning (integrating new classes while retaining old knowledge) under strict memory and computational constraints.

2. Methodology

The proposed framework operates as a unified pipeline consisting of four main stages:

A. Semantic Embedding Learning

Input Processing: Raw I/Q RF signals are transformed into time-frequency spectrograms using Short-Time Fourier Transform (STFT) and normalized.
Hybrid Encoder: A dual-branch architecture extracts features:
- CNN Branch: Uses multiscale dilated convolutions to capture fine-grained spectral textures and wideband temporal envelopes.
- Transformer Branch: Uses self-attention with sinusoidal positional encodings to model long-range dependencies and preserve time-frequency locality.
Feature Fusion: Intermediate features are fused via Multi-Nonlinear Layers (MNL) into a compact semantic embedding ( $z$ ).
Loss Function: Trained using a composite loss ( $L_{feat}$ ) combining intra-class compactness, inter-class margin separation, and cross-entropy to ensure distinct clustering of known classes.

B. Open-Set Recognition (OSR)

Mechanism: Known classes are modeled using Gaussian statistics (mean $\mu_k$ and covariance $\Sigma_k$ ) in the embedding space.
Decision Rule: During inference, the Mahalanobis distance ( $d_k$ ) between a test sample and known class distributions is calculated.
Thresholding: A dynamic threshold ( $\tau_k$ $τ_{k}$ ) is derived using the three-sigma rule (99.7% confidence boundary).
- If $d_k < \tau_k$ : The sample is classified as the known class $k$ .
- If $d_k \ge \tau_k$ : The sample is rejected as "unknown" and stored in a buffer ( $U$ ) for further processing.

C. Clustering for Novel Class Discovery

Preprocessing: Rejected samples are standardized. If dimensionality is high (>64), Principal Component Analysis (PCA) is applied to reduce dimensions and mitigate the curse of dimensionality.
Algorithm Selection: The system evaluates both K-Means and Gaussian Mixture Models (GMM).
Optimal Cluster Selection: A Composite Validity Score ( $Q(k)$ ) is calculated to select the best clustering result. This score integrates:
- Silhouette Score ( $S_k$ )
- Calinski-Harabasz Index ( $CH_k$ )
- Davies-Bouldin Index ( $DB_k$ )
- Explained Variance ( $V_k$ )
- Elbow Method: Used to detect inflection points in inertia curves.
Filtering: Clusters with high purity and appropriate size are retained as new UAV categories ( $C_{new}$ ).

D. Incremental Learning with Lightweight Replay

Integration: Newly discovered classes are integrated into the classifier without full retraining.
Replay Mechanism: A memory-bounded replay buffer ( $M$ ) stores a limited number of representative samples from old classes ( $C_{old}$ ).
Training: The model is updated using a mixed batch of new samples and replayed old samples. This minimizes the cross-entropy loss ( $L_{CE}$ ) while preserving the decision boundaries of previously learned classes, effectively mitigating catastrophic forgetting.

3. Key Contributions

Unified Framework: A scalable pipeline integrating OSR, clustering-based discovery, and replay-based IL for RF-based UAV recognition.
Robust Clustering Strategy: An automatic model-order selection mechanism using a composite validity score to ensure reliable discovery of new classes even in noisy RF environments.
Memory-Efficient Adaptation: A lightweight incremental learning mechanism that uses bounded replay to maintain stability (retention of old knowledge) and plasticity (learning new classes) without large-scale retraining.

4. Experimental Results

The framework was evaluated on a real-world dataset containing 24 UAV classes (18 known, 6 unknown) across 915 MHz, 2.4 GHz, and 5.8 GHz bands.

Open-Set Detection: The Mahalanobis-based decision boundary effectively separated unknown signals from known classes, preventing overconfident misclassifications.
Clustering Performance:
- The proposed composite selection strategy automatically chose K-Means over GMM.
- Purity: Achieved 92.1% purity for discovered clusters, significantly outperforming GMM (90.6%).
- Visual Validation: t-SNE visualizations showed strong alignment between predicted clusters and ground-truth unknown classes, though some fragmentation occurred due to intra-class variability (e.g., different flight modes).
Incremental Learning:
- Without Replay: Catastrophic forgetting occurred, dropping old class accuracy to 0%.
- With Minimal Replay: A small replay budget (5 samples per old class) restored old class accuracy to 100% while maintaining ~98% accuracy on new classes.
- Increasing the replay size beyond this minimal threshold yielded negligible improvements, proving the efficiency of the approach.

5. Significance and Future Work

Significance: This work addresses the "open-world" challenge in UAV surveillance, moving beyond static, closed-set models. It demonstrates that reliable, lifelong learning for RF signals is possible with minimal memory overhead, making it suitable for field-deployable systems with limited storage.
Limitations & Future Directions:
- Current evaluation uses a fixed class split; future work will test random unknown-class selections.
- The Gaussian approximation for OSR thresholds may need refinement for highly complex RF environments.
- Future research will explore domain-invariant feature extraction, real-time embedded deployment, and federated learning across distributed monitoring nodes.